> Suno on the other hand to me has basically been a complete joke. The training data is just not wide enough to do anything interesting and new.

Hard disagree, Suno is terrific, but likely depends on what you define as "new". Of course AI models can't explore outside the data they are trained on, but they can effectively generate within the parameter space, including areas unexplored directly by the content it was trained on.

Have done some really interesting mashups of styles that people aren't doing and the results are great imho.

If you are trying to make it do very specific things like you would another instrument, it's not the right tool.